Prediction of ICU Readmissions using LSTM in Low and Middle Income Countries

Abstract

ABSTRACT This study addresses the persistent challenge of Intensive Care Unit (ICU) readmission, focusing on the unique context of Lower and Middle Income Countries (LMICs). Despite advancements in medical technology, ICU readmissions remain a critical issue, with implications for healthcare resources and patient outcomes. To address the challenge of ICU readmission, accurate prediction models are needed to identify patients at high risk of readmission because the prediction of readmission before the patient is discharged, will help physicians re-evaluate the discharge of the patient and reduce the immature discharges. The existing literature predominantly stems from high-income countries (HICs), and this study aims to fill the gap by developing a predictive model tailored to LMICs context. It utilizes the Long Short-Term Memory (LSTM), known for its ability to capture temporal dependencies in sequential patients’ data to predict the early ICU readmission (readmission within 48 hours followed by index discharge) of the patients and feature ablation test to extract the important factors associated with ICU readmission. 2.85% (306) of discharges to the wards were later readmitted within 48 hours to the intensive care unit. The LSTM model with a cost-sensitive training had significantly better performance (area under the receiver operating curve, 0.68) compared to the baseline models with traditional machine learning approaches. It highlights that the deep learning models improve the accuracy of decision-making in predicting ICU readmission.